Kuaishou Search Ads Deploys New Generative Retrieval Paradigm: UniGD Integrates Generation and Discrimination to Reshape Retrieval Pipeline

Kuaishou Technology's research team proposed a unified generation-discrimination framework called UniGD, which jointly optimizes SID generation and explicit relevance discrimination in a shared model. It uses semantic codebooks to construct stable, immediately available content representations for ads and preserves dedicated semantic spaces for different material types. During online inference, UniGD first generates candidate SIDs and then reuses internal model representations for lightweight relevance scoring, replacing the original cascaded external relevance model. Experiments show that this design improves both generative recall and relevance discrimination accuracy while reducing the latency of the entire pipeline.